Comment by vjerancrnjak 21 hours ago
If it overfits on the whole internet then it’s like a search engine that returns really relevant results with some lossy side effect.
Recent benchmark on unseen 2025 Math Olympiad shows none of the models can problem solve . They all accidentally or on purpose had prior solutions in the training set.
The claim is that these models are training on data which include the problems and explanations. The fact that the first model trained after the public release of the questions (and crowdsourced answers) performs best is not a counter example, but is expected and supported by the claim.
I was noodling with Gemini 2.5 Pro a couple days ago and it was convinced Donald Trump didn’t win the 2024 election and that he conceded to Kamala Harris so I’m not entirely sure how much weight I’d put behind it.
You probably mean the USAMO 2025 paper. They updated their comparison with Gemini 2.5 Pro, which did get a nontrivial score. That Gemini version was released five days after USAMO, so while it's not entirely impossible for the data to be in its training set, it would seem kind of unlikely.
https://x.com/mbalunovic/status/1907436704790651166